Liqid unveils single server reference architecture with 16 GPUs

Liqid launched a reference architecture at SC23 to build a composable system with a Dell PowerEdge server fitted with up to 16 Nvidia GPUs.

The UltraStack L40S scheme is based on the PowerEdge R760xa fitted with Nvidia L40S GPUs. The R760xa is designed for mainstream AI workloads and at launch could support up to four double-width GPUs and up to 12 single-width GPUs. Liqid says access to high-end GPUs, like Nvidia’s H100 Tensor Core GPU, have higher lead times, and many customers need eight or more GPUs in a single server.

Sumit Puri, Liqid CEO and co-founder, said: “Our Liqid UltraStack reference architecture is a game-changer, built around the formidable Nvidia L40S GPUs and Dell Technologies PowerEdge servers. It’s not just a concept – it’s a tangible, reliable solution readily available. This innovation provides an unparalleled advantage for compute-intensive tasks, firmly placing you ahead of the curve.”

Liqid UltraStack

Liqid’s technology dynamically composes virtual servers using disaggregated pools of separate CPU+DRAM, GPU, Optane SSDs, FPGA, networking, and storage resources accessed over a PCIe fabric. The composed servers are torn down when no longer needed and their elements returned to the source pools for reuse. Liqid can compose memory, meaning DRAM and SCM like Optane DIMMs, separately from CPU+DRAM and so achieve a degree of external memory pooling.

Now Liqid claims the UltraStack reference architecture delivers the world’s first 16-way L40S system, combining L40S GPUs, ConnectX-7 NICs, BlueField DPUs, and Liqid NVMe IO Accelerator cards in a single Dell R760xa server. It claims this fundamentally redefines the limits of high-density GPU systems and is needed for generative AI, HPC, and graphics-intensive workloads.

The L40S has a fabric-integrated chassis and a qualified composable system, with a single SKU for Liqid software, hardware, and support. It includes an AI stack that’s community supported and open source. Liqid says the system is simple to buy, configure, and operate. unlike Nvidia’s H100 (Hopper) and A100 GPUs, the L40S with its Ada Lovelace architecture does not suffer from long lead times.

Liqid UltraStack

Liqid claims that, compared to using a 4 x 2RU servers, each containing 4 x L40S GPUs, its 16 x L40S GPU UltraStack system provides 35 percent more performance, a 75 percent reduction in software licensing, 35 percent reduction in electrical energy, and 25 percent lower total cost of ownership.

Liqid UltraStack

 There are x8 and x16 L40S variants of the UltraStack reference architecture. Download an UltraStack L40S datasheet here.